[Hematoma Segmentation of Spontaneous Intracerebral Hemorrhage Based on Watershed and Region-Growing Algorithm].

Objective: To establish a brain hematoma CT image segmentation method based on watershed and region-growing algorithm so as to measure hematoma volume quickly and accurately, to explore the consistency between the results of this segmentation method and those of manual segmentation, the clinical gold standard, and to compare the results of this method with the calculation of the two Tada formulas commonly used in clinical practice. Methods: The preoperative CT images of 152 patients who were treated for spontaneous cerebral hemorrhage at the Department of Neurosurgery, West China Hospital, Sichuan University between January 2018 and June 2019 were retrospectively collected. The CT images were randomly assigned, by using a random number table, to the training set, the test set and the validation set, which contained 100 patients, 22 patients and 30 patients, respectively. The labeling results of the training set and the test set were used in algorithm training and testing. Four methods, namely, manual segmentation, algorithm segmentation, i.e., segmentation calculation based on watershed and regional growth algorithm, Tada formula, i.e., the traditional Tada formula calculation, and accurate Tada formula, i.e., accurate Tada formula calculation based on 3D-Slicer, were applied on the validation set to measure the hematoma volume. The Digital Imaging and Communications in Medicine (DICOM) data of subjects meeting the selection criteria of the study were manually segmented by two experienced neurosurgeons. The hematoma segmentation model was built based on watershed algorithm and regional growth algorithm. Seed point selected by neurosurgeons was taken as the starting point of growth. Regional grayscale difference criterion combined with manual segmentation validation were adopted to determine the regional growth threshold that met the segmentation precision requirements for intracranial hematoma. Using manual segmentation as the gold standard, Bland-Altman consistency analysis was used to verify the consistency of the three other methods for measuring hematoma volume. Results: With manual segmentation as the gold standard, among the three methods of measuring hematoma volume, algorithm segmentation had the smallest percentage error, the narrowest range of difference, the highest intra-group correlation coefficient (0.987), good consistency, and the narrowest 95% limits of agreement ( LoA). The percentage error of its segmentation was not statistically significant for hematomas of different volumes. Conclusion: The segmentation method of spontaneous intracerebral hemorrhage based on watershed and regional growth algorithm shows stable measurement performance and good consistency with the clinical gold standard, which has considerable clinical significance, but it still needs further validation with more clinical samples.

[Application of Improved Unet Network in the Recognition and Segmentation of Hemorrhage Regions in Brain CT Images].

OBJECTIVE: To examine the performance and application value of improved Unet network technology in the recognition and segmentation of hemorrhage regions in brain CT images. METHODS: A total of 476 brain CT images of patients with spontaneous intracerebral hemorrhage (SICH) were retrospectively included. The improved Unet network was used to identify and segment the hemorrhage regions in the patients' brain CT images. The CT imaging data of the hemorrhage regions were manually labelled by clinicians. After randomized sorting, 430 data sets from 106 patients were selected for inclusion in the training set and 46 data sets from 11 patients were included in the test set. After data enhancement, the experimental data set underwent network training and model testing in order to assess the segmentation performance. The segmentation results were compared with the those of the Unet network (Base), FCN-8s network and Unet++ network. RESULTS: In the segmentation of brain CT image hemorrhage region with the improved Unet network, the three evaluation indicators of Dice similarity coefficient, positive predictive value (PPV), and sensitivity coefficient (SC) reached 0.8738, 0.9011 and 0.8648, respectively, increasing by 8.80%, 7.14% and 8.96%, respectively, compared with those of FCN-8s, and increasing by 4.56%, 4.44% and 4.15%, respectively, compared with those of Unet network (Base). The improved Unet network also showed better segmentation performance than that of Unet++ network. CONCLUSION: The improved method based on Unet network proposed in this report displayed good performance in the recognition and segmentation of hemorrhage regions in brain CT images, and is an appropriate method for the recognition and segmentation of hemorrhage regions in brain CT images, showing potential application value for assisting clinical decision-making and preventing early hematoma expansion.